AI for Fisheries Science: Neural Network Tools for Forecasting, Spatial Standardization, and Policy Optimization

14 The development of Artificial Intelligence (AI) presents novel opportunities for tackling 15 complex marine resource management challenges. Among AI models, neural networks are a 16 powerful class of tools capable of learning nonlocal and lagged patterns from fisheries data as 17 well as approximating nonlinear relationships among multiple latent variables using estimation 18

that automatically implement statistical shrinkage. This gives them potential to 19 effectively handle data obtained from fisheries populations subject to dynamic environments. We 20 highlight two flexible subclasses and one application of neural networks: Long Short-Term 21 Memory (LSTM) and Convolutional Neural networks (CNNs) and policy discovery via 22 Reinforcement Learning. LSTMs are designed to handle sequential data by allowing prediction 23 from past values at both short and long time-lags. CNNs are not explicitly designed to handle 24 temporal information, but can interpolate a spatial latent variable based on its value within a 25 geographic neighborhood, and can be combined with LSTM models for this purpose. This "Food 26 for Thought" paper introduces and applies these neural network approaches, both alone and in 27 combination, to demonstrate their potential application for several foundational topics in 28 fisheries science: 1) the forecasting of population weight-at-age, 2) the standardization of spatio-29 temporal indices of relative abundance, and 3) the discovery of harvest policies to optimize 30 catches and maintain spawning biomass. Each section provides a simple, simulated example and 31 describes the tradeoffs – particularly the lack of inferential capability – presented by using neural 32 networks over traditional approaches for each topic. We then outline medium-term research 33 questions that may clarify, facilitate or qualify the applicability of these tools to fisheries 34 management science. Finally, we discuss how future combinations of these approaches could 35

in simplified ways to estimate and forecast stock biomass and provide harvest advice. 36 37 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 1

38 Motivation 39 Science based fisheries management is currently facing unprecedented challenges due to shifting 40 funding priorities and dynamic and unexpected changes in the marine environment (Patrick and 41 Link, 2015). These challenges are co-occurring with an expansion of methodologies designed to 42 address the diverse data types now available for some fisheries, from molecular information 43 informing movement and demographic estimation (Punt et al., 2024) to fine-scale acoustic 44 monitoring data (Griffin et al., 2018). However, fisheries data have traditionally been analyzed 45 using biological models fit statistically to data, which are constrained by, potentially biased, 46 model assumptions. Artificial intelligence (AI) tools are less constraining in their assumptions 47 and have been explored for data collection and preliminary analysis (such as the statistical post-48 stratification of fisheries data (Gasper and Cahalan, 2025, p. 202), occurrence records (Morand et 49 al., 2024) or automated image detection (Saqib et al., 2024), but the broader potential of AI 50 remains underexplored. There are simultaneous efforts to modernize the modelling infrastructure 51 used to assess fishery populations (Maunder et al., 2025), presenting an opportune moment to 52 revisit the methodological landscape of fisheries stock assessment. This paper focuses on neural 53 network models as a promising avenue to modernize scientific fisheries management in the 54 coming decades. We provide readers the core concepts of neural network models, in contrast to 55 familiar concepts from generalized linear modeling and machine learning and share three 56 demonstrative case studies. 57 Pointwise Regression: A Common Tool in Fisheries Management 58 Most existing fisheries management modeling approaches rely on regression and parametric 59 process based models. These include traditional approaches such as generalized linear models 60 (GLMs) through many machine learning (ML) methods, which include boosted regression trees, 61 random forests, projection pursuit regression, lasso, and Gaussian process models (to name a 62 few; Hastie et al., 2001), and has been used in fisheries science for over a decade (Rubbens et al., 63 2023). ML and GLMs are similar in that they both typically define a pointwise regression, 64 where each sample is predicted by a vector of associated covariates; these can include complex 65

of handling spatio-temporal processes, such as tinyVAST (Thorson et al., 2025). Due 66 to this conceptual similarity between ML and GLMs, there have been many previous analyses in 67 fisheries science comparing ML methods with conventional GLMs or hierarchical models (Stock 68 et al., 2020), or extending regression models to include a Gaussian process component (Sugeno 69 and Munch, 2013; Thorson et al., 2014). These methods have the advantage of being easy to 70 implement and facilitate inference, but are limited by assumptions regarding the statistical 71 distribution of modeled data and are sensitive to mis-specification. 72 Neural Networks: an overview of a foundational AI technique 73 Neural networks are a class of machine learning models inspired by the structure and function of 74 the mammalian brain. They are constructed using a series of transformations of features and 75 penalized model weights that can approximate any function. This architecture offers advantages 76 for handling high-dimensional, nonlinear and spatiotemporal data types common in fisheries; 77 their value to ecology for this reason has been recognized for nearly three decades (Lek et al., 78 1996). In some cases, neural networks can outperform traditional process-based or statistical 79 models in predicting ecological processes. Alternatively, neural networks can be embedded 80 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 2 within process based models to improve performance (e.g., (Triebe et al., 2021; Wesselkamp et 81 al., 2024)). 82 83 We claim in the following that neural networks (NNs) are often more suited than previous ML 84

to identify complex features that are present in fisheries analysis because of their ability 85 to learn nonlinear and complex patterns. Neural networks penalize complexity differently from 86 GLMs by the progressive updating of model weights (Fan et al., 2021), and can optionally set a 87 subset of model weights to zero to prevent overfitting (Srivastava et al., 2014). These approaches 88 are known as “implicit” and “explicit” regularization, respectively. Usefully, these methods for 89 regularization do not require marginalizing across any coefficients, and are therefore much faster 90 than regularization in Bayesian or empirical-Bayes hierarchical models. Current NN methods 91 have also seen less use in fisheries science; one aim of the present paper is to promote the 92 exploration of these methods. Table 1 provides an explicit comparison between familiar concepts 93 in statistical-ecology and equivalent (though not always identical) concepts in neural 94 networks/artificial intelligence 95 Potential applications of Neural Networks in Fisheries 96 There are numerous types of neural network models, and this paper focuses on two for their 97 potential suitability to the spatial and/or temporal dynamism common to marine fisheries. 98 Recurrent Neural Networks (RNNs) are a subtype of neural network designed to handle 99 sequential data such as time series using feedback loops that maintain an internal state or 100 memory of previous inputs in the series. This recurrent structure enables RNNs to capture 101 temporal dependencies and model the evolution of dynamical systems. They are natural 102 candidates for nonlinear autoregressive models, where future values are predicted based on past 103 values. 104 105 However, basic RNNs can suffer from the numerical underflow or overflow issues (see Table 1) 106 during training, which makes it difficult for them to learn long-range dependencies in the data. 107 Long Short-Term Memory (LSTM) networks are a specialized type of RNN architecture 108 designed to mitigate these issues. LSTMs introduce a cell state, which acts as a long-term 109 memory, and sub-states known as “gates” that control the flow of information into and out of the 110 cell state, allowing the network to selectively remember and forget information over long 111 sequences. The value of RNNs like LSTMs lies in their ability to model and predict the behavior 112 of partially observed dynamical systems, where not all relevant state variables are directly 113 measured. By learning the temporal patterns in the observed data, RNNs can make predictions 114 about future states, which is highly valuable in fisheries science for forecasting population 115 dynamics, catch, and other time-dependent variables. 116 117 Convolutional Neural Networks (CNNs) were developed for the computer vision field as a 118 technique to facilitate the detection and labeling of images. In the simplest case, this involves a 119 multi-step process of passing (“convolving”) a set of learnable convolutional filters, or kernels, 120 to iteratively extract patterns. This process builds a hierarchical numerical representation, often 121 referred to as a feature map or embedding, that captures high-level features at low spatial 122 resolution. This allows the CNN to effectively learn spatial dependencies and patterns in the 123 image, which could be a photograph of an animal or a map of observed biomass from a fishery 124 independent survey, presenting an intriguing possibility for the standardization of spatially-125 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 3 explicit data. However, the basic CNN framework is not designed to explicitly handle temporal 126 data nor irregular or sparse datasets characteristic of most fisheries surveys or catch time series. 127 Researchers have combined CNN and LSTM neural network approaches to produce forecasts of 128 spatial processes (e.g., Yang et al., (2025); the authors are aware of a single example wherein 129 CNN and LSTM neural network approaches were combined to produce forecasts of probable 130 catches (Agmata and Guðmundsson, 2025), although other studies have used CNN in isolation 131 (Morand et al., 2024). Crucially, that example did not conduct an explicit comparison between 132 the proposed CNN+LSTM approach, variations thereof, and currently-used methods for handling 133 spatio-temporal data, such as design-based expansion or regression-based standardization tools 134 such as tinyVAST, which we do in this case study. 135 Aims and Structure of this Paper 136 This "Food for Thought" paper introduces and applies these neural network approaches, both 137 alone and in combination, as a forward-looking demonstration for several foundational topics in 138 fisheries science: 1) the forecasting of population processes, with size-at-age as an example; 2) 139 the standardization of spatio-temporal indices of relative abundance, and 3) the discovery of 140 harvest policies to optimize catches and maintain spawning biomass. For each case study we 141 describe the methods used to produce the illustrative example, and discuss the relevance for the 142 fisheries assessment and management audience. Where possible, we provide direct mapping 143 between existing concepts in statistical ecology (such as spatio-temporal standardization or 144 interpolation) or fisheries management (such as management strategy evaluation, MSE; for this 145 reason, the section on reinforcement learning for management is longer than the others). 146 Advancing these techniques could enhance the adaptability, precision and sustainability of 147 scientific fisheries management in a rapidly changing environment. The paper concludes with a 148 “Research Roadmap” outlining the most important research questions to be considered in 149 evaluating the tractability of these tools for fisheries management in the coming decades. 150 151 Case Studies 152 This paper presents three case studies. Each is designed to highlight how NNs may be applied to 153 a variety of data types and processes central to fisheries management science. We also selected 154 these to represent a spectrum of tractability and impact, ranging from applications that could be 155 implemented today to those that would require completely novel management frameworks 156 Figure 1. We emphasize that the case study methods are examples and are not final; more work 157 is needed by the fisheries science community to investigate the tradeoffs, specifications, and 158 nuances of each application before they are ready for operational use. 159 160 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 1 Case Study 1: Neural Networks for Stock Demography Forecasting 161 Even highly complex stock assessment models typically resort to simplification of observed 162 demographic processes in order to make management decisions. In particular, projections of 163 future size-at-age form the basis for fisheries management advice, yet scientists will commonly 164 use a historical or recent (e.g. 5-year) mean of observed fish sizes, though size-at-age can change 165 dramatically among years (Stawitz et al., 2019). Stock assessment models also rely on estimates 166 of historical size-at-age to derive estimates of historical biomass. Process-based models, such as 167 the von-Bertalanffy growth curve, are commonly used for estimating average size-at-age. Given 168 that size-at-age is well-observed for many managed fisheries, is sensitive to stochastic 169 environmental processes, and underpins catch advice for industrial fisheries management, this 170 case study illustrates how NN can provide flexibility to predict observable processes in fisheries. 171 172 We specifically compared the performance of two NN approaches to several existing methods 173 for deriving size-at-age. The models included a simple average of size-at-age from the terminal 174 five years of data (‘mean-5’), a three parameter von-Bertalanffy curve (‘VBGF’); a three 175 parameter von-Bertalanffy curve with IID year univariate random effects on asymptotic size and 176 growth rate (VBGF-RE’); a von Bertalanffy curve with 3D AR1 year, age, and cohort random 177 effects (‘GMRF’; Cheng et al., 2023); a simple 3-layer NN (‘NN’), and a LSTM (‘LSTM’); see 178 Supplementary Material S.1. for further details.. 179 180 To explore whether the relative performance of the different approaches depend on the temporal 181 structure of the simulation, we fit all six models to two simulated size-at-age datasets with 182 parameters based on Bering Sea pollock (Gadus chalcogrammus) from NOAA’s Alaska 183 Fisheries Science Center. The first simulated dataset simulated data from a three parameter von-184 Bertalanffy curve representing constant time-invariant growth. The second simulated dataset 185 simulated data from a three parameter von-Bertalanffy curve with time-varying parameters 186 following an AR1 and directional trend representing time-varying growth. Model parameters and 187 sample sizes used for simulated data-sets were based on sampling effort and life history 188 parameters of cod like species. We produced 300 30-year replicates of each dataset. Only 189 simulations in which all models converged were retained. 190 191 We evaluated prediction performance on each dataset by calculating the average root mean 192 squared error (RMSE) across ages from a 10-fold cross validation where random years of data 193 were removed from each fold. One- and two-year projection performance was evaluated using 194 five retrospective forecast peels. This allowed us to compare average RMSE across peels, both 195 for hindcast and projection accuracy and between the time-invariant or time-dependent scenarios. 196 197 This simulation exercise demonstrated that the LSTM most frequently resulted in the lowest 198 RMSE for all three prediction periods (hindcast, one and two years forward) for the time-199 invariant growth simulations (Figure 2). When simulated size-at-age varied through time, 200 average RMSE was most frequently lowest for the GMRF model for the hindcast and one-year 201 projection (Figure 2), though the LSTM and mean-5 performed comparably or better for the 202 two-year projection in this scenario. 203 204 The GMRF is designed to handle temporal variation explicitly and is able to separate observation 205 uncertainty from the latent temporal variability; it is possible that with lower observation error 206 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 2 the performance of the LSTM method would have been improved. For both size-at-age 207 scenarios, we also visualized how predictive performance varied across ages, which had a less 208 distinct pattern (Supplementary Figures S1 and S2). 209 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 1 Case Study 2: Standardizing survey data through space and time using a convolutional 210 neural network 211 Several methods have been developed to model or standardize for spatio-temporal processes in 212 fisheries data, particularly survey observations. These methods arose from the recognition that 213 underlying population processes and the methods for observing fish populations are subject to 214 variation in space and time, and accounting for this variation is required to ensure unbiased 215 estimates of population size or trajectory. The tinyVAST framework (Thorson et al., 2025) 216 builds upon the Vector Autoregressive Spatio-Temporal (VAST, Thorson (2019)) modeling 217 approach that allows users to specify separable Gaussian Markov Random fields and delta-218 GLMMs. The latest approach allows the specification of a broader class and flexibility of 219 multivariate models including spatial factor analysis and ARIMA (Jenkins, 2014) structures, 220 enabling the representation of simultaneous, lagged and recursive dependencies common to 221 ecological and fishery processes. TinyVAST is similar to sdmTMB (Anderson et al., 2022) in 222 that they both utilize the modern Template Model Builder language (Kristensen et al., 2016), 223 incorporate SPDE-based spatial precision matrices, though the former is better suited for 224 allowing for multivariate temporal dependencies. Surprisingly, the recognition that neural 225 network models are useful for data with strong spatial dependencies (Wikle and Zammit-226 Mangion, 2023) has not yet led to rigorous investigation of how these techniques compare to 227 existing approaches for standardizing biomass observations into annual indices of fishery 228 abundance. 229 230 The simulation study was conducted on a 20 x 20 spatial grid over 12 years, with spatial 231 correlation modeled through a row-standardized neighborhood matrix (rho = 0.95) and temporal 232 correlation through an AR1 process (p = 0.8). The underlying simulated biomass included both 233 spatial and temporal random effects; in each year 200 cells (50% of the domain) were randomly 234 sampled for observation that arose from a Tweedie distribution (Zainol Mustafa and Nadia, 235 2025; rho = 1.5, psi = 0.2). The experiment was replicated 25 times to assess model performance 236 and stability. For each replicate, observed data were passed to two CNN-based models 237 (described below) as well as to tinyVAST to a) interpolate continuous spatio-temporal biomass 238 estimates and b) calculate annual indices of abundance. A design-based estimator was included 239 for comparison for the annual indices. Design-based expansion is a simple calculation that takes 240 the average observation across sampled cells and multiplies it across the entire domain based on 241 the fraction of observed cells; this precludes the production of continuous surfaces but facilitates 242 comparison of annual indices. 243 244 We included two versions of the CNN approach, which differed in their handling of missing data 245 and temporal dependencies. Both approaches relied on the network to learn spatial relationships 246 through coordinate embeddings and produced continuous surfaces of estimated biomass. The 247 simpler approach, labeled “CNN” in results, took spatial coordinates as input and processed each 248 year separately; only points with observed (sampled) data were included in training. These 249 vectorized inputs passed through two dense embedding layers (dimensions 64 and 128, with 250 swish activation), reshaped to a 16x4x2 spatial tensor and processed by two 3x3 two-dimensional 251 convolutional layers with ReLU activation. The second approach, labeled “CNN-LSTM” 252 explicitly handled sparsity in the observed data via a binary masking channel which informed the 253 model where observations existed in a given year. Temporal information in the CNN-LSTM was 254 incorporated via a lag-based approach, whereby predictions at each timestep were informed by 255 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 2 the previous three years of data. Both CNN models were implemented in R using the keras3 256 package (Kalinowski et al., 2025), compiled with Adam and a trainable Tweedie loss function 257 with sigmoid power parameter initialized to 1.5 and constrained to (1, 2) specified to support 258 comparison with tinyVAST. The 12-year, 25-repliate experiment to fit the three models took 16 259 hours to run on a personal computer. Performance was compared across models by examining 260 the trajectory of standardized indices, and overall RMSE in log-biomass and log-density (across 261 all replicates and years). 262 263 This simulation exercise demonstrated that neither of the CNN-based models out-performed the 264 existing tinyVAST approach in terms of log-biomass and log-density, though all approaches 265 were able to produced standardized indices of similar scale to the true biomass (Figure 3). recent 266 work has replicated our finding that CNNs can match, but not surpass, traditional species 267 distribution modeling approaches (Kellenberger et al., 2026). Examination of the interpolated 268 map for a single year and replicate shows that while both CNN-based models were able to 269 identify areas of relatively higher and lower biomass only tinyVAST was able to recover small-270 scale regional patterns. This application of CNNs certainly extends beyond the traditional 271 implementation, wherein an intact image (or collection of images for training) would be passed 272 to the model. CNNs are known to perform poorly when images are incompletely observed 273 (Heinke et al., 2021), and are also known to perform poorly when training data are limited (most 274 CNNs are trained on thousands of images, not 12 years of data as in this case) 275 276 There are promising applications of neural networks that could be used for the detection of 277 spatial patterns, some of which are in active development (e.g., (Deng et al., 2025)) and would 278 likely require the generation of continuous spatio-temporal loadings from observed data. 279 However, this case study indicates that tinyVAST is an appropriately precise and efficient 280

for the straightforward task of index standardization using sparse data; more 281 sophisticated extensions to the CNN approach such as a vision transformer (Dosovitskiy et al., 282 2021) could be revisited for well-sampled fishery populations, and/or from interpolated meshes 283 derived from such data. 284 285 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 1 Case Study 3: Setting next year’s catch using reinforcement learning (RL) 286 A common problem is fisheries is the selection of management strategies to achieve policy 287 objectives. Management strategy evaluation is a simulation framework that was developed to 288 evaluate alternative strategies under uncertainty. Traditional MSE is a closed-loop forward-in-289 time simulation framework used to evaluate the performance of harvest strategies under 290 uncertainty that can include an operating model (OM), estimation model (EM) and management 291 rule. MSE is inherently sequential and forward-looking, as it mimics real-world decision-making 292 over time. Crucially, the mechanism (or ‘policy’) by which future catches are determined is 293 nearly always a predefined rule which is not informed or modified by the simulation itself. This 294 means it is not tractable to explore all possible policies using traditional MSE. MSE is a time-295 consuming process that requires stakeholder input (Goethel et al., 2019) and the manual 296 specification of alternative states of nature, estimation models, and harvest policies (Punt et al., 297 2016). Recent work has highlighted that static reference points like B0 (unfished biomass, the 298 foundation of many management systems) are highly sensitive to model assumptions (Edgar et 299 al., 2024) and fundamentally problematic for data-poor or dynamic fisheries to the degree that 300 empirical data streams may offer more robust guidance for management in such situations 301 (Blamey et al., 2025). These issues justify the exploration of data-driven methods, such as 302 reinforcement learning, that can discover policies autonomously. 303 304 Reinforcement learning is a subfield of machine learning that is dedicated to solving sequential 305 decision-making problems and consists of two primary components: the agent and the 306 environment. The RL agent is a trainable neural network that interacts repeatedly with the 307 ‘environment’, a representation of the population dynamics akin to the operating model used in 308 Management Strategy Evaluation (Butterworth, 2007; Punt et al., 2016). During this interactive 309 process (called “training”) the agent receives a feedback signal known as the “reward1” which it 310 seeks to optimize by updating a policy, analogous to a management rule in MSE. A key 311 distinction between MSE and RL is that the policy in RL is updated according to the agent’s 312 experience in the environment, whereas the HCR in MSE remains static across simulations. 313 314 This case study directly compared the performance of a traditional management strategy 315 evaluation with a RL-derived policy for the Eastern Bering Sea (EBS) pollock fishery. This 316 allowed us to investigate the tractability and relative performance of the RL approach for a 317 fishery with complex dynamics (multiple fishing fleets, age-structure) and a high degree of data 318 availability (including weight-at-age and compositional data), yet using a simple, singular 319

(maximization of annual harvest). This case study therefore extends beyond recent 320 contributions that focus on simplified or simulated stock dynamics (e.g., Montealegre-Mora et al. 321 (2025)). Two recruitment scenarios were explored for both approaches: the first utilized a stock 322 independent approach where expected recruitment was defined by random annual deviations 323 around a constant value (‘Constant recruitment’), and the second utilized a Ricker stock-324 recruitment relationship, where stock output varied both following random annual deviations and 325 as a function of stock size (‘Ricker recruitment’). For EBS pollock, recruitment dynamics are a 326 noted uncertainty for this stock and both parameterizations have been explored for management 327 advice (Ianelli et al., 2024). This allowed us to determine whether commonly-used recruitment 328 dynamics impacted relative performance or stability of the RL approach for this fishery. We 329 1 For readers seeking a broader introduction to RL in the context of ecology, see Lapeyrolerie et al., (2022). 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 2 compared future performance of the catch recommendations determined by the MSE and RL 330 approaches by examining the median simulation trajectory and uncertainty of stock spawning 331 biomass realized catches during the projection period. For illustration, we also compared the 332 harvest policy learned by the RL approach to the survey observations and realized catches from 333 the MSE. Full specifications of this case study can be found in Supplementary Material section 334 2. 335 336 The reinforcement learning (RL) agent trained under constant recruitment developed a harvest 337 policy resembling precautionary harvest control rules. It avoided fishing when bottom trawl 338 survey biomass was below about 10,500 mt and increased harvest sharply to ~4 million mt above 339 that threshold, reaching ~4.25 million mt at higher levels. This produced large year-to-year 340 fluctuations—ranging from zero to nearly double historical catches—but consistently maintained 341 spawning biomass above that of the management strategy evaluation (MSE) scenario, which had 342 stable catches (~1.5 million mt) and biomass (~2.5 million mt).Under the Ricker recruitment 343 scenario, the RL agent’s policy was non-linear but smoother than in the constant recruitment 344 case. It maintained nearly constant catches (1.8–3 million mt) for survey observations above 345 4,000 t, unexpectedly increased catches around 2,500 t, and stopped fishing below 1,000 t. This 346 strategy stabilized spawning biomass near 1.25 million mt (~21% of unfished levels; 20%B0 is a 347 common cutoff for fishery closure) while sustaining harvests of ~3.5 million mt—substantially 348 higher than the MSE scenario, which averaged 1–1.5 million mt of catch and 2–2.5 million mt of 349 spawning biomass. Despite its unconventional shapes, the RL policy outperformed the MSE in 350 cumulative yield while maintaining more stable trajectories under the more complex Ricker 351 dynamics. 352 353 These results expand upon findings by Montealegre-Mora (2025) to suggest that the RL 354 approach is capable of discovering harvest policies for complex, data-rich fisheries and produce 355 catch and spawning biomass trajectories of similar magnitude to those obtained by MSE under 356 two distinct recruitment scenarios. That study demonstrated that RL can provide insights into 357 HCR design that conventionally used methods in fisheries management are unable to achieve. 358 This is especially intriguing given that the RL did not need to “step through” an updated 359 estimation model at each timestep, presenting a potential companion or alternative to the time-360 consuming process of updating stock assessments in resource-limited settings. 361 362 Future work could explore constraints on the policy curve (i.e., forcing recti-linear policies or 363 setting strict upper limits on catches, though these can inhibit agent training). The RL method 364 could also be extended to base the current state of the population on more data types (age 365 composition data, for instance) or to include representations of observation or process error (e.g., 366 via curriculum learning, see Table 2). Finally, it would be useful to explore comparison of the 367 RL method to empirical harvest control rules that set catches based on recent survey 368 observations, particularly in the context of data quality and availability. The RL method 369 employed here assumes that annual survey observations are equally representative of the 370 population among years and through time; it would be useful to know if there are lower limits to 371 the frequency and precision of survey observations that facilitate RL learning. 372 373 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 3 Summary of findings from Case Studies 374 Our case studies highlight the flexibility that neural networks present in addressing foundational 375 topics in fisheries science, their immediate shortcomings, and directions for future research. For 376 the prediction of weight-at-age, it appears valuable to include LSTM-based approaches as part of 377 an analytical pipeline to evaluate forecast and hindcast projection accuracy given the frequency 378 with which they out-performed other methods in terms of RMSE. For some prediction 379 categories, the GMRF method showed similar performance in the presence of temporal variation 380 in growth. For the development of annual indices of abundance from spatio-temporal survey 381 data, or interpolation of biomass in space, neither the CNN nor CNN-LSTM approaches 382 appeared to outperform tinyVAST. It is important to recognize that the CNN technique was 383 developed for computer-vision applications under the assumption that the entire image is 384 available to the network; the sparse information provided by our simulated survey was 385 insufficient for the CNN to accurately characterize spatial dependencies, is precisely what 386 tinyVAST was designed to handle. The RL-to-MSE comparison exercise highlighted promising 387 potential for novel policy discovery, with the tradeoff that absent strong guidance, policies might 388 be un-intuitive (hindering stakeholder communication and support) or induce untenably large 389 variation in population trajectories. 390 391 Food for Thought: Priority Research Questions for AI in Fisheries Science 392 When statistical catch-at-age software tools such as Stock Synthesis became widespread, the 393 fisheries science community spent nearly two decades examining the impacts of 394 ‘misspecification’ on model performance, highlighting how un-accounted-for dynamics in spatial 395 structure, fish growth, selectivity patterns, and more could bias the estimation of management 396 quantities and subsequent advice. Now the fisheries science community must apply that same 397 focus to uncovering the trade-offs and biases presented by using neural networks in tandem with 398 our existing process-based modeling workflows. 399 400 Here, we pose several outstanding questions regarding the potential use of neural network 401 models for fisheries science, which we categorize based on their use in projections, for setting 402 quotas, or for fisheries science and research (Table 3). In particular, fisheries process models 403 were designed to represent the mechanisms that underly population responses to fishing 404 (Beverton and Holt, 1957). By representing these mechanisms, fisheries scientists presumably 405 hoped to develop a “structural causal model” (Pearl, 2009), which could then be used to predict 406 fishery responses to policy decisions that have not previously been seen (i.e., the simultaneous 407 impact of fishing and climate change). We therefore recommend specific comparison of existing 408 process-based and new neural-network models regarding their ability to predict (and quantify 409 uncertainty) about fishery responses to previously unobserved policies; fishery scientists may 410 need to expand their vocabulary for describing sources of uncertainty to include those common 411 in the AI literature (for example, Allen Akselrud, 2024). Our first two case studies emphasized 412 “predictive performance,” and good predictive performance is no guarantee of good performance 413 for disentangling multiple causes when recommending new policies (Arif and MacNeil, 2022). 414 Given the large literature regarding causal modelling in artificial intelligence research (reviewed 415 briefly in Luo et al., 2020), we are optimistic that future neural network research can develop 416 robust fisheries policies using monitoring data for systems involving multiple drivers. 417 418 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 4

419 This Food-for-thought paper sought to 1) introduce the scientific fishery management 420 community to modern techniques in deep learning with neural networks, exploring specific 421 applications to foundational fisheries topics. Our case studies highlight that these methods have 422 promise to supplement, improve upon, or change the ways we manage industrial fisheries, and 423 underscore that rigorous benchmarking (comparison to existing methods) should be employed as 424 they are further explored and refined. 425 426 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 5

427 428 Agmata, A., Guðmundsson, S., 2025. Convolutional-LSTM approach for temporal catch 429 hotspots (CATCH): an AI-driven model for spatiotemporal forecasting of fisheries catch 430 probability densities. Biol. Methods Protoc. 10, bpaf045. 431 https://doi.org/10.1093/biomethods/bpaf045 432 Allen Akselrud, C.I., 2024. Random forest regression models in ecology: Accounting for messy 433 biological data and producing predictions with uncertainty. Fish. Res. 280, 107161. 434 https://doi.org/10.1016/j.fishres.2024.107161 435 Anderson, S.C., Ward, E.J., English, P.A., Barnett, L.A.K., Thorson, J.T., 2022. sdmTMB: An R 436 Package for Fast, Flexible, and User-Friendly Generalized Linear Mixed Effects Models 437 with Spatial and Spatiotemporal Random Fields. 438 https://doi.org/10.1101/2022.03.24.485545 439 Arif, S., MacNeil, M.A., 2022. Predictive models aren’t for causal inference. Ecol. Lett. 25, 1741–440 1745. https://doi.org/10.1111/ele.14033 441 Beverton, R.J.H., Holt, S.J., 1957. On the dynamics of exploited fish populations. Chapman & 442 Hall, London. 443 Blamey, L.K., Plagányi, É.E., Moore, T., 2025. Exploring dynamic reference points for a short-444 lived, highly variable species under future climate. Can. J. Fish. Aquat. Sci. 82, 1–16. 445 https://doi.org/10.1139/cjfas-2025-0070 446 Butterworth, D.S., 2007. Why a management procedure approach? Some positives and 447 negatives. ICES J. Mar. Sci. 64, 613–617. https://doi.org/10.1093/icesjms/fsm003 448 Deng, R., Li, Z., Wang, M., 2025. GeoAggregator: An Efficient Transformer Model for Geo-449 Spatial Tabular Data, in: The Thirty-Ninth AAAI Conference on Artificial Intelligence 450 (AAAI-25). Presented at the Association for the Advancement of Artificial Intelligence, 451 Association for the Advancement of Artificial Intelligence, pp. 11572–11580. 452 Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, 453 M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N., 2021. An Image is 454 Worth 16x16 Words: Transformers for Image Recognition at Scale. 455 https://doi.org/10.48550/arXiv.2010.11929 456 Edgar, G.J., Bates, A.E., Krueck, N.C., Baker, S.C., Stuart-Smith, R.D., Brown, C.J., 2024. 457 Stock assessment models overstate sustainability of the world’s fisheries. Science 385, 458 860–865. https://doi.org/10.1126/science.adl6282 459 Fan, J., Ma, C., Zhong, Y., 2021. A Selective Overview of Deep Learning. Stat. Sci. 36, 264–460 290. https://doi.org/10.1214/20-STS783 461 Gasper, J., Cahalan, J., 2025. Utilizing the random forest algorithm and interpretable machine 462 learning to inform post-stratification of commercial fisheries data. Fish. Res. 281, 463 107253. https://doi.org/10.1016/j.fishres.2024.107253 464 Goethel, D.R., Lucey, S.M., Berger, A.M., Gaichas, S.K., Karp, M.A., Lynch, P.D., Walter, J.F., 465 Deroba, J.J., Miller, S., Wilberg, M.J., 2019. Closing the feedback loop: on stakeholder 466 participation in management strategy evaluation. Press. Addressing Fish. Chall. Manag. 467 Strategy Eval. 01, 1895–1913. https://doi.org/10.1139/cjfas-2018-0162@cjfas-468 mse.issue01 469 Griffin, L.P., Brownscombe, J.W., Adams, A.J., Boucek, R.E., Finn, J.T., Heithaus, M.R., 470 Rehage, J.S., Cooke, S.J., Danylchuk, A.J., 2018. Keeping up with the Silver King: 471 Using cooperative acoustic telemetry networks to quantify the movements of Atlantic 472 tarpon (Megalops atlanticus) in the coastal waters of the southeastern United States. 473 Fish. Res. 205, 65–76. https://doi.org/10.1016/j.fishres.2018.04.008 474 Hastie, T., Tibshirani, R., Friedman, J., 2001. The Elements of Statistical Learning, Springer 475 Series in Statistics. Springer New York Inc., New York, NY, USA. 476 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 6 Heinke, D., Wachman, P., van Zoest, W., Leek, E.C., 2021. A failure to learn object shape 477 geometry: Implications for convolutional neural networks as plausible models of 478 biological vision. Vision Res. 189, 81–92. https://doi.org/10.1016/j.visres.2021.09.004 479 Jenkins, G.M., 2014. Autoregressive–Integrated Moving Average (ARIMA) Models, in: Wiley 480 StatsRef: Statistics Reference Online. John Wiley & Sons, Ltd. 481 https://doi.org/10.1002/9781118445112.stat03472 482 Kalinowski, T., Allaire, J.J., Chollet, F., 2025. keras3: R Interface to “Keras.” 483 Kellenberger, B., Winner, K., Jetz, W., 2026. The Performance and Potential of Deep Learning 484 for Predicting Species Distributions. Glob. Ecol. Biogeogr. 35, e70184. 485 https://doi.org/10.1111/geb.70184 486 Kristensen, K., Nielsen, A., Berg, C.W., Skaug, H., Bell, B.M., 2016. TMB: Automatic 487 differentiation and laplace approximation. J. Stat. Softw. 488 https://doi.org/10.18637/jss.v070.i05 489 Lapeyrolerie, M., Chapman, M.S., Norman, K.E.A., Boettiger, C., 2022. Deep reinforcement 490 learning for conservation decisions. Methods Ecol. Evol. 13, 2649–2662. 491 https://doi.org/10.1111/2041-210X.13954 492 Lek, S., Delacoste, M., Baran, P., Dimopoulos, I., Lauga, J., Aulagnier, S., 1996. Application of 493 neural networks to modelling nonlinear relationships in ecology. Ecol. Model. 90, 39–52. 494 https://doi.org/10.1016/0304-3800(95)00142-5 495 Luo, Y., Peng, J., Ma, J., 2020. When causal inference meets deep learning. Nat. Mach. Intell. 496 2, 426–427. https://doi.org/10.1038/s42256-020-0218-x 497 Maunder, M.N., Crone, P.R., Semmens, B.X., Valero, J.L., Waterhouse, L., Methot, R.D., Punt, 498 A.E., 2025. The Center for the Advancement of Population Assessment Methodology 499 (CAPAM): A perspective on the first 10 years. Fish. Res. 281, 107162. 500 https://doi.org/10.1016/j.fishres.2024.107162 501 Montealegre-Mora, F., Boettiger, C., Walters, C.J., Cahill, C.L., 2025. Using Machine Learning 502 to Inform Harvest Control Rule Design in Complex Fishery Settings. Fish Fish. 26, 1004–503 1020. https://doi.org/10.1111/faf.70013 504 Morand, G., Joly, A., Rouyer, T., Lorieul, T., Barde, J., 2024. Predicting species distributions in 505 the open ocean with convolutional neural networks. Peer Community J. 4. 506 https://doi.org/10.24072/pcjournal.471 507 Patrick, W.S., Link, J.S., 2015. Myths that Continue to Impede Progress in Ecosystem‐ Based 508 Fisheries Management. Fisheries 40, 155–160. 509 https://doi.org/10.1080/03632415.2015.1024308 510 Pearl, J., 2009. Causality. Cambridge University Press, Cambridge, UK. 511 Punt, A.E., Butterworth, D.S., de Moor, C.L., De Oliveira, J.A.A., Haddon, M., 2016. 512 Management strategy evaluation: Best practices. Fish Fish. 17, 303–334. 513 https://doi.org/10.1111/faf.12104 514 Punt, A.E., Thomson, R., Little, L.R., Bessell-Browne, P., Burch, P., Bravington, M., 2024. 515 Including close-kin mark-recapture data in statistical catch-at-age stock assessments 516 and management strategies. Fish. Res. 276, 107057. 517 https://doi.org/10.1016/j.fishres.2024.107057 518 Rubbens, P., Brodie, S., Cordier, T., Destro Barcellos, D., Devos, P., Fernandes-Salvador, J.A., 519 Fincham, J.I., Gomes, A., Handegard, N.O., Howell, K., Jamet, C., Kartveit, K.H., 520 Moustahfid, H., Parcerisas, C., Politikos, D., Sauzède, R., Sokolova, M., Uusitalo, L., 521 Van den Bulcke, L., van Helmond, A.T.M., Watson, J.T., Welch, H., Beltran-Perez, O., 522 Chaffron, S., Greenberg, D.S., Kühn, B., Kiko, R., Lo, M., Lopes, R.M., Möller, K.O., 523 Michaels, W., Pala, A., Romagnan, J.-B., Schuchert, P., Seydi, V., Villasante, S., Malde, 524 K., Irisson, J.-O., 2023. Machine learning in marine ecology: an overview of techniques 525 and applications. ICES J. Mar. Sci. 80, 1829–1853. 526 https://doi.org/10.1093/icesjms/fsad100 527 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 7 Saqib, M., Khokher, M.R., Yuan, X., Yan, B., Bearham, D., Devine, C., Untiedt, C., Cannard, T., 528 Maguire, K., Tuck, G.N., Little, L.R., Wang, D., 2024. Fishing event detection and 529 species classification using computer vision and artificial intelligence for electronic 530 monitoring. Fish. Res. 280, 107141. https://doi.org/10.1016/j.fishres.2024.107141 531 Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., Salakhutdinov, R., 2014. Dropout: a 532 simple way to prevent neural networks from overfitting. J Mach Learn Res 15, 1929–533 1958. 534 Stawitz, C.C., Haltuch, M.A., Johnson, K.F., 2019. How does growth misspecification affect 535 management advice derived from an integrated fisheries stock assessment model? Fish. 536 Res. 213, 12–21. https://doi.org/10.1016/j.fishres.2019.01.004 537 Stock, B.C., Ward, E.J., Eguchi, T., Jannot, J.E., Thorson, J.T., Feist, B.E., Semmens, B.X., 538 2020. Comparing predictions of fisheries bycatch using multiple spatiotemporal species 539 distribution model frameworks. Can. J. Fish. Aquat. Sci. 77, 146–163. 540 https://doi.org/10.1139/cjfas-2018-0281 541 Sugeno, M., Munch, S.B., 2013. A semiparametric Bayesian approach to estimating maximum 542 reproductive rates at low population sizes. Ecol. Appl. 23, 699–709. 543 https://doi.org/10.1890/12-0453.1 544 Thorson, J.T., 2019. Guidance for decisions using the Vector Autoregressive Spatio-Temporal 545 (VAST) package in stock, ecosystem, habitat and climate assessments. Fish. Res. 210, 546 143–161. https://doi.org/10.1016/j.fishres.2018.10.013 547 Thorson, J.T., Anderson, S.C., Goddard, P., Rooper, C.N., 2025. tinyVAST: R Package With an 548 Expressive Interface to Specify Lagged and Simultaneous Effects in Multivariate Spatio-549 Temporal Models. Glob. Ecol. Biogeogr. 34, e70035. https://doi.org/10.1111/geb.70035 550 Thorson, J.T., Ono, K., Munch, S.B., 2014. A Bayesian approach to identifying and 551 compensating for model misspecification in population models. Ecology 95, 329–341. 552 https://doi.org/10.1890/13-0187.1 553 Triebe, O., Hewamalage, H., Pilyugina, P., Laptev, N., Bergmeir, C., Rajagopal, R., 2021. 554 NeuralProphet: Explainable Forecasting at Scale. 555 https://doi.org/10.48550/arXiv.2111.15397 556 Wesselkamp, M., Moser, N., Kalweit, M., Boedecker, J., Dormann, C.F., 2024. Process-557 Informed Neural Networks: A Hybrid Modelling Approach to Improve Predictive 558 Performance and Inference of Neural Networks in Ecology and Beyond. Ecol. Lett. 27, 559 e70012. https://doi.org/10.1111/ele.70012 560 Wikle, C.K., Zammit-Mangion, A., 2023. Statistical Deep Learning for Spatial and 561 Spatiotemporal Data. Annu. Rev. Stat. Its Appl. 10, 247–270. 562 https://doi.org/10.1146/annurev-statistics-033021-112628 563 Yang, H., Chen, C., Xue, X., 2025. Coupled convolutional neural network with long short-term 564 memory network for predicting lake water temperature. J. Hydrol. 655, 132878. 565 https://doi.org/10.1016/j.jhydrol.2025.132878 566 Zainol Mustafa, Nadia, 2025. Tweedie Distribution: A Statistical Solution for Unusually 567 Dispersed Data. Sci. J. Stat. Probab. Its Appl. 3, 29–37. 568 https://doi.org/10.24127/sciencestatistics.v3i1.8003 569 570 571 572 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 8 Tables 573 574 Term in statistical ecology Term in neural networks Brief description deterministic latent variable layer A hidden but non-random state predicted from the model nonlinear transformation for element of a latent variable neuron A method of abstraction to map unobserved effects to responses, for example, through logistic or exponential curves reverse-mode automatic differentiation backpropagation Algorithm to efficiently compute gradients by working backwards through the model parameter weight A model value controlling the strength of an input's effect parameter estimation training The process of adjusting parameters or weights to minimize error on observed data regression modelling supervised learning Predicting outcomes from inputs using observed data gradient-based optimizer2 Stochastic gradient descent (e.g., Adam) Methodology for updating model parameters during training to minimize error optimizing a management policy stochastic policy ascent A process of adjusting policy parameters to optimize a specific reward function numerical overflow and underflow Exploding or vanishing gradient Because the gradient calculation involves products of terms, and each term in the product might be very large or small, the gradient might 2 Where each optimization step is based on a subset of data; hyperparameters control how the explored parameters update given the gradient, and each loop through all partitions of the data is called an “epoch” 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 9 end up being smaller or larger than the numerical minimum or maximum allowed. This can preclude accurate calculation of the gradient. 575 Table 1. Non-exhaustive correspondence of vocabulary from statistical ecology to neural 576 network/AI methods. 577 578 579 Reinforcement Learning for Fisheries Management Strategy Evaluation Representation of population/system dynamics Operating model conditioned to historical data, termed “environment” Operating model conditioned to historical data Data source used for model selection Future simulation or historical data Always future simulation Representation of future stochasticity Simulated variation in future recruitment deviations Simulated variation in future population processes, typically recruitment deviations Representation of observation and/or estimation uncertainty Can be addressed with specialized neural network architectures (i.e., environmental uncertainty as hidden state) Handled explicitly by estimation model How catches are selected The policy function, which is learned from direct interactions with the environment, prescribes the annual catch according to the observed state Pre-defined harvest control rule sets annual catches based on the current state (which may be one or more estimated quantities or observed data) How performance is measured reward function is maximized during training as the agent interacts with the environment; performance metrics can be calculated after training based on the operating model time series, After simulations, calculate performance metrics using operating model time series during projection period 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 10 as in MSE Comparison across policies Harvest policies compared internally as part of learning. Novel harvest control rules explored and evaluated; can consider distinct yet flexible rules; no consideration of estimation models Can consider distinct yet pre- defined harvest control rules and/or estimation models as strategy components Table 2. Comparison of key concepts in reinforcement learning-based policy discovery and 580 Management Strategy Evaluation. 581 582 583 For use in projections and management modeling ● Does including neural-network derived projections of weight-at-age improve management performance? Can these approaches be extended to other key population dynamics (i.e., numbers at age of recruitment?) ● Neural networks are not natively designed to handle nor represent uncertainty in their predictions, which is something highly valued by the fishery management community and often a requirement of reporting. How should we represent uncertainty in projection models that use neural networks to project one or more components? ● If a neural network can predict data almost perfectly, how closely must that data represent the system for the predictions to be acceptable for management? ● Is it possible to construct a neural network that predicts stock biomass using a feature array of input data more accurately than a process-based estimation model? ● If so, can such a model be used to explore the impacts of varying catch levels on future population biomass (and thus fully replace the process-based projection framework used to calculate management quantities)? ● Can neural networks support the simulation component of management strategy evaluation, perhaps enhancing the accuracy of predicted population dynamics, observed data, or both? For use in quota setting ● Can the data predicted by a neural network replace true observational data for short- term management use (for example, to inform an empirical harvest control rule in the absence of a survey), and if so, how long is “short term”? ● How might a management system integrate processed-based models with RL-derived policies? At what frequency would operating models/environments need to be updated to confidently rely upon a policy learned by RL? ● How would the development of RL-based policies be communicated to the fishery community, and how would stakeholder engagement change under this paradigm? For the fisheries science community ● Are there historical, social or economic characteristics of a fishery and management system that are not well suited for these applications? ● What are the ethical considerations of providing management advice based on AI models that inherently offer less explainability than existing methods? 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 11 Table 3: List of questions arising from case studies, that could form the basis of future research 584 in scientific fisheries management. 585 586 587 588 589 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 12 Figures 590 591 592 593 Figure 1. Conceptual diagram of case studies. The case studies were selected and are ordered 594 (from left to right) in terms of how tractably each NN application could be included in current 595 scientific fishery management frameworks. 596 597 598 Figure 2. Case Study 1: Number of times each process- and neural network based model resulted 599 in the lowest average root mean squared error (RMSE) when fit to simulated data (n = 300) 600 without (top row) or with (bottom row) a temporal trend in true fish growth. Hindcast represents 601 10-fold leave year out cross-validation and “y+1” and “y+2” represents forecast skill from 602 sequential peels of historical data and forecasted for two future years. 603 12 ed 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 13 604 Figure 3. Case Study 2: Investigation of neural networks for spatio-temporal survey data 605 interpolation and index standardization. A) Mean scaled annual indices (lines) and 95% 606 confidence interval (shaded ribbon) for 25 replicates of a simulated 12-year biomass time series; 607 black solid line is true biomass whereas all other colors are model estimates. B) Performance 608 metrics of various estimation models (colors) in terms of RMSE in biomass (lefthand figure) or 609 density (righthand side) for estimates in a 20x20 gridded domain for 12 simulated years across 610 25 replicates. C) Maps of true and estimated log biomass across the domain for a single replicate 611 and year; results shown only for models that produce an interpolated surface as part of the 612 estimation step. 613 614 13 s; te 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 14 615 Figure 4. Case Study 3: Investigation of neural-network based reinforcement learning method 616 (orange lines and points) for fisheries management, in comparison to 25 replicates of an MSE 617 (blue lines and points) for EBS pollock using an OM conditioned on observations from 1964-618 2024 (grey rectangle; years are compressed) with a projection period from 2025-2050. Top row: 619 realized catch (millions of mt) under the constant (1a) and Ricker (1b) recruitment scenarios. 620 Middle row: stock spawning biomass in the operating model under the constant (2a) and Ricker 621 (2b) recruitment scenarios. Horizontal dashed line indicates 20% of unfished biomass and the 622 colored line represents the median across simulation replicates for a given year. Colored ribbons 623 represent 95% simulation interval. Bottom row: realized harvest policy in terms of the quota 624 (millions of mt) specified for future years(s) (points) versus the observed Bottom Trawl Survey 625 Biomass (t) arising from each method under the constant (3a) or Ricker (3b) recruitment 626 scenarios. The MSE did not use a survey-based method for setting quotas, but observations from 627 25 MSE replicates and the subsequent years’ catch are shown for comparison. 628 629 14 m 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint 105 and is also made available for use under a CC0 license. (which was not certified by peer review) is the author/funder. This article is a US Government work. It is not subject to copyright under 17 USC The copyright holder for this preprintthis version posted March 17, 2026. ; https://doi.org/10.64898/2026.03.13.711664doi: bioRxiv preprint